Resting-EEG Characteristics and Classification Model Exploration for Tap Test Prediction in Communicating
Nan Bai1, Zijia Guo2, Wenxuan Xin2
1Department of Neurosurgery, Nanjing Jinling Hospital, Affiliated Hospital of Medical School, Nanjing University, 210093 Nanjing, Jiangsu, China.
Background:
Communicating hydrocephalus is characterized by gait disturbance, cognitive impairment, and altered consciousness. The cerebrospinal fluid (CSF) tap test is widely used to predict shunt responsiveness; however, objective neurophysiological markers are still lacking. Resting-state electroencephalography (RS-EEG) may offer insights into brain network function.
Methods:
Thirty patients with communicating hydrocephalus underwent 19‑channel RS‑EEG before the CSF tap test. Multimodal EEG features were extracted from 2‑second epochs, including spectral power, functional connectivity (coherence and phase‑locking value), and nonlinear measures (approximate entropy, Hjorth parameters), then aggregated to subject‑level representations. Group differences were explored via correlation with clinical scales, source‑level phase transfer entropy, and t-distributed Stochastic Neighbor Embedding (t‑SNE) visualization. For classification, logistic regression, linear support vector machine (SVM), and a deep neural network (fully connected or 1D Convolutional Neural Network [1D‑CNN]) were trained and compared under nested leave‑one‑subject‑out cross‑validation with within‑fold feature selection. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), permutation tests, bootstrap 95% confidence intervals, and DeLong tests.
Results:
Responders showed higher alpha-band power, which correlated with Coma Recovery Scale-Revised (CRS-R) improvement (r = -0.764, p = 0.0004) but not Mini-Mental State Examination (MMSE). Functional connectivity and source-level analyses showed descriptive trends that did not survive false discovery rate correction. t-SNE visualization showed partial separation of responders and non-responders, and EEG data quality did not differ between groups. For classification, linear models matched or exceeded deep learning: in the disorders of consciousness (DOC) subgroup, logistic regression (LR) and SVM achieved an AUC of 0.982 (95% CI: 0.909-1.000) & 0.982 (95% CI: 0.893-1.000) vs. 0.964 (0.84-1.00) for the fully connected network; in the cognitive subgroup, SVM reached an AUC of 0.963 (0.84-1.00), followed by CNN (0.944) and LR (0.926). All models were significant by permutation test (p < 0.002). DeLong tests showed no significant difference between deep learning and the best linear model (p > 0.4), indicating that EEG features are largely linearly separable. Bootstrap confidence intervals remained above 0.73.
Conclusion:
These findings indicate that EEG-derived spectral features, particularly alpha and theta power, contain linearly separable information related to tap test responsiveness. Linear machine learning models performed on par with deep learning, underscoring the robustness of the neural signal and cautioning against unnecessary model complexity in small clinical samples. This exploratory analysis demonstrates the feasibility of EEG-based prediction but requires prospective external validation.
